We are extending our solution for one of the the outbound channel and planning to implement delayed learning for adaptive model.
Now the requirement is for outbound channel, we will wait for a response for 48 hours and if we don't receive any response then we will update the model with impression.
I saw couple of configuration in Pega CDH
1. "Response timeout" in Prediction studio where i can mention the time beyond which alternate level(impression in our case) will be updated.
2. "Make decision and store data for later response capture" in Dataflow->strategy configuration, where we can provide the time till which data will be stored for response update.
1. Are these 2 time configuration(mention in point 1 & 2) should be same? if yes, then why pega kept it open for configuration separately?
2. In case in 1st configuration point(response timeout) I give 24 hr. and in the 2nd configuration point i give 48 hr. to store the data for later response capture then will system update twice the adaptive model one with alternate level and then with actual response?
The two settings have very different scope as well as function. The first is at prediction level and the other is for the system. So what you are telling the system with the second setting is, keep decision data around fir delayed learning for a max of X across all predictions. After this amount of time the data is purged. This setting does not activate any delayed learning within the window of time specified.
Now for the prediction setting, certain predictions, say web predictions, you will wait less time before assuming assuming a customer has not responded and so you will use this as a trigger to process negative impressions. For another prediction, you might wait a little longer. So for this setting, it is defining a prediction level window and it’s also defining a processing activity (process negative impressions). There is no purging of data within this timeframe, it’s just sent to adaptive for modeling.
Given these differences, you should set the second setting to the widest time window for predictions it supports. So if you have 2 inbound predictions with 2 different response processing windows, make sure the system wide setting is the larger of those 2 settings, probably with a little bit of a buffer.